Download Report

Design of Parallel and High-Performance
Computing
Fall 2013
Lecture: Locks and Lock-Free
Instructor: Torsten Hoefler & Markus Püschel
TA: Timo Schneider
Administrivia

Next week – progress presentations
 Make sure, Timo knows about your team (this step is important!)
 Send slides (ppt or pdf) by Sunday 11:59pm to Timo!
 10 minutes per team (hard limit)
 Prepare! This is your first impression, gather feedback from us!
 Rough guidelines:
Present your plan
Related work (what exists, literature review!)
Preliminary results (not necessarily)
Main goal is to gather feedback, so present some details
Pick one presenter (make sure to switch for other presentations!)

Intermediate (very short) presentation: Thursday 11/21 during
recitation

Final project presentation: Monday 12/16 during last lecture
2
Review of last lecture

Language memory models
 History
 Java/C++ overview

Locks
 Two-thread
 Peterson
 N-thread
 Many different locks, strengths and weaknesses
 Lock options and parameters

Formal proof methods
 Correctness (mutual exclusion as condition)
 Progress
3
DPHPC Overview
4
Goals of this lecture

Hardware operations for concurrency control

More on locks (using advanced operations)
 Spin locks
 Various optimized locks

Even more on locks (issues and extended concepts)
 Deadlocks, priority inversion, competitive spinning,
semaphores

Case studies
 Barrier
 Reasoning about semantics

Locks in practice: a set structure
5
Lamport’s Bakery Algorithm (1974)

Is a FIFO lock (and thus fair)

Each thread takes number in doorway and threads enter in the order
of their number!
volatile int flag*n+ = ,0,0,…,0-;
volatile int label*n+ = ,0,0,….,0-;
void lock() {
flag[tid] = 1; // request
label[tid] = max(label[0], ...,label[n-1]) + 1; // take ticket
while ((∃k != tid)(flag[k] && (label[k],k) <* (label[tid],tid))) {};
}
public void unlock() {
flag[tid] = 0;
}
6
Lamport’s Bakery Algorithm

Advantages:
 Elegant and correct solution
 Starvation free, even FIFO fairness

Not used in practice!
 Why?
 Needs to read/write N memory locations for synchronizing N threads
 Can we do better?
Using only atomic registers/memory
7
A Lower Bound to Memory Complexity

Theorem 5.1 in [1]: “If S is a [atomic] read/write system with at least
two processes and S solves mutual exclusion with global progress
[deadlock-freedom], then S must have at least as many variables as
processes”

So we’re doomed! Optimal locks are available and they’re
fundamentally non-scalable. Or not?
[1] J. E. Burns and N. A. Lynch. Bounds on shared memory for mutual
exclusion. Information and Computation, 107(2):171–184, December
1993
8
Hardware Support?

Hardware atomic operations:
 Test&Set





Write const to memory while returning the old value
Atomic swap
Atomically exchange memory and register
Fetch&Op
Get value and apply operation to memory location
Compare&Swap
Compare two values and swap memory with register if equal
Load-linked/Store-Conditional LL/SC
Loads value from memory, allows operations, commits only if no other updates
committed  mini-TM
Intel TSX (transactional synchronization extensions)
Hardware-TM (roll your own atomic operations)
9
Relative Power of Synchronization

Design-Problem I: Multi-core Processor
 Which atomic operations are useful?

Design-Problem II: Complex Application
 What atomic should I use?

Concept of “consensus number” C if a primitive can be used to solve the
“consensus problem” in a finite number of steps (even if a threads stop)
 atomic registers have C=1 (thus locks have C=1!)
 TAS, Swap, Fetch&Op have C=2
 CAS, LL/SC, TM have C=∞
10
Test-and-Set Locks


Test-and-Set semantics
 Memoize old value
 Set fixed value TASval (true)
 Return old value
After execution:
 Post-condition is a fixed (constant) value!
bool test_and_set (bool *flag) {
bool old = *flag;
*flag = true;
return old;
} // all atomic!
11
Test-and-Set Locks

Assume TASval indicates “locked”

Write something else to indicate “unlocked”

TAS until return value is != TASval

When will the lock be
granted?
volatile int lock = 0;

Does this work well in
practice?
void lock() {
while (TestAndSet(&lock) == 1);
}
void unlock() {
lock = 0;
}
12
Contention


On x86, the XCHG instruction is used to implement TAS
 For experts: x86 LOCK is superfluous!
Cacheline is read and written
 Ends up in exclusive state, invalidates other copies
 Cacheline is “thrown” around uselessly
 High load on memory subsystem
movl $1, %eax
xchg %eax, (%ebx)
x86 bus lock is essentially a full memory barrier 
13
Test-and-Test-and-Set (TATAS) Locks

Spinning in TAS is not a good idea

Spin on cache line in shared state
 All threads at the same time, no cache coherency/memory traffic

Danger!
 Efficient but use with great
care!
 Generalizations are
dangerous
volatile int lock = 0;
void lock() {
do {
while (lock == 1);
} while (TestAndSet(&lock) == 1);
}
void unlock() {
lock = 0;
}
14
Warning: Even Experts get it wrong!

Example: Double-Checked Locking
1997
Problem: Memory ordering leads to race-conditions!
15
Contention?

Do TATAS locks still have contention?

When lock is released, k threads fight for
cache line ownership
 One gets the lock, all get the CL exclusively (serially!)
 What would be a good
solution? (think “collision
avoidance”)
volatile int lock = 0;
void lock() {
do {
while (lock == 1);
} while (TestAndSet(&lock) == 1);
}
void unlock() {
lock = 0;
}
16
TAS Lock with Exponential Backoff

Exponential backoff eliminates contention statistically
 Locks granted in
unpredictable
order
 Starvation possible
but unlikely
How can we make
it even less likely?
volatile int lock = 0;
void lock() {
while (TestAndSet(&lock) == 1) {
wait(time);
time *= 2; // double waiting time
}
}
void unlock() {
lock = 0;
}
Similar to: T. Anderson: “The performance of spin lock alternatives for shared-memory
multiprocessors”, TPDS, Vol. 1 Issue 1, Jan 1990
17
TAS Lock with Exponential Backoff

Exponential backoff eliminates contention statistically
 Locks granted in
unpredictable
order
 Starvation possible
but unlikely
Maximum waiting
time makes it less
likely
volatile int lock = 0;
const int maxtime=1000;
void lock() {
while (TestAndSet(&lock) == 1) {
wait(time);
time = min(time * 2, maxtime);
}
}
void unlock() {
lock = 0;
}
Similar to: T. Anderson: “The performance of spin lock alternatives for shared-memory
multiprocessors”, TPDS, Vol. 1 Issue 1, Jan 1990
18
Comparison of TAS Locks
19
Improvements?

Are TAS locks perfect?
 What are the two biggest issues?
 Cache coherency traffic (contending on same location with expensive
atomics)
-- or - Critical section underutilization (waiting for backoff times will delay entry
to CR)

What would be a fix for that?
 How is this solved at airports and shops (often at least)?

Queue locks -- Threads enqueue
 Learn from predecessor if it’s their turn
 Each threads spins at a different location
 FIFO fairness
20
Array Queue Lock


Array to implement
queue
 Tail-pointer shows next free
volatile int array*n+ = ,1,0,…,0-;
volatile int index*n+ = ,0,0,…,0-;
volatile int tail = 0;
queue position
 Each thread spins on own
location
void lock() {
CL padding!
index[tid] = GetAndInc(tail) % n;
 index[] array can be put in TLS while (!array[index[tid]]); // wait to receive lock
}
So are we done now?
 What’s wrong?
 Synchronizing M objects
requires Θ(NM) storage
 What do we do now?
void unlock() {
array[index[tid]] = 0; // I release my lock
array[(index[tid] + 1) % n] = 1; // next one
}
21
CLH Lock (1993)

List-based (same queue
principle)

Discovered twice by Craig,
Landin, Hagersten 1993/94


2N+3M words
 N threads, M locks
Requires thread-local qnode
pointer
 Can be hidden!
typedef struct qnode {
struct qnode *prev;
int succ_blocked;
} qnode;
qnode *lck = new qnode; // node owned by lock
void lock(qnode *lck, qnode *qn) {
qn->succ_blocked = 1;
qn->prev = FetchAndSet(lck, qn);
while (qn->prev->succ_blocked);
}
void unlock(qnode **qn) {
qnode *pred = (*qn)->prev;
(*qn)->succ_blocked = 0;
*qn = pred;
}
22
CLH Lock (1993)

Qnode objects represent
thread state!
 succ_blocked == 1 if waiting
or acquired lock
 succ_blocked == 0 if released
lock

List is implicit!
 One node per thread
 Spin location changes
NUMA issues (cacheless)

Can we do better?
typedef struct qnode {
struct qnode *prev;
int succ_blocked;
} qnode;
qnode *lck = new qnode; // node owned by lock
void lock(qnode *lck, qnode *qn) {
qn->succ_blocked = 1;
qn->prev = FetchAndSet(lck, qn);
while (qn->prev->succ_blocked);
}
void unlock(qnode **qn) {
qnode *pred = (*qn)->prev;
(*qn)->succ_blocked = 0;
*qn = pred;
}
23
MCS Lock (1991)

Make queue explicit
 Acquire lock by
appending to queue
 Spin on own node
until locked is reset

Similar advantages
as CLH but
 Only 2N + M words
 Spinning position is fixed!
Benefits cache-less NUMA

What are the issues?
 Releasing lock spins
 More atomics!
typedef struct qnode {
struct qnode *next;
int succ_blocked;
} qnode;
qnode *lck = NULL;
void lock(qnode *lck, qnode *qn) {
qn->next = NULL;
qnode *pred = FetchAndSet(lck, qn);
if(pred != NULL) {
qn->locked = 1;
pred->next = qn;
while(qn->locked);
}}
void unlock(qnode * lck, qnode *qn) {
if(qn->next == NULL) , // if we’re the last waiter
if(CAS(lck, qn, NULL)) return;
while(qn->next == NULL); // wait for pred arrival
}
qn->next->locked = 0; // free next waiter
qn->next = NULL;
}
24
Lessons Learned!

Key Lesson:
 Reducing memory (coherency) traffic is most important!
 Not always straight-forward (need to reason about CL states)

MCS: 2006 Dijkstra Prize in distributed computing
 “an outstanding paper on the principles of distributed computing, whose
significance and impact on the theory and/or practice of distributed
computing has been evident for at least a decade”
 “probably the most influential practical mutual exclusion algorithm ever”
 “vastly superior to all previous mutual exclusion algorithms”
 fast, fair, scalable  widely used, always compared against!
25
Time to Declare Victory?

Down to memory complexity of 2N+M
 Probably close to optimal

Only local spinning
 Several variants with low expected contention

But: we assumed sequential consistency 
 Reality causes trouble sometimes
 Sprinkling memory fences may harm performance
 Open research on minimally-synching algorithms!
Come and talk to me if you’re interested
26
More Practical Optimizations

Let’s step back to “data race”
 (recap) two operations A and B on the same memory cause a data race if
one of them is a write (“conflicting access”) and neither AB nor BA
 So we put conflicting accesses into a CR and lock it!
This also guarantees memory consistency in C++/Java!

Let’s say you implement a web-based encyclopedia
 Consider the “average two accesses” – do they conflict?
27
Reader-Writer Locks

Allows multiple concurrent reads
 Multiple reader locks concurrently in CR
 Guarantees mutual exclusion between writer and writer locks and reader
and writer locks

Syntax:
 read_(un)lock()
 write_(un)lock()
28
A Simple RW Lock


Seems efficient!?
 Is it? What’s wrong?
 Polling CAS!
Is it fair?
 Readers are preferred!
 Can always delay
writers (again and
again and again)
const W = 1;
const R = 2;
volatile int lock=0; // LSB is writer flag!
void read_lock(lock_t lock) {
AtomicAdd(lock, R);
while(lock & W);
}
void write_lock(lock_t lock) {
while(!CAS(lock, 0, W));
}
void read_unlock(lock_t lock) {
AtomicAdd(lock, -R);
}
void write_unlock(lock_t lock) {
AtomicAdd(lock, -W);
}
29
Fixing those Issues?

Polling issue:
 Combine with MCS lock idea of queue polling

Fairness:
 Count readers and writers
(1991)
The final algorithm (Alg. 4)
has a flaw that was
corrected in 2003!
30
Deadlocks

Kansas state legislature: “When two trains approach each other at a
crossing, both shall come to a full stop and neither shall start up again
until the other has gone.”
[according to Botkin, Harlow "A Treasury of Railroad Folklore" (pp. 381)]
What are necessary
conditions for deadlock?
31
Deadlocks

Necessary conditions:
 Mutual Exclusion
 Hold one resource, request another
 No preemption
 Circular wait in dependency graph

One condition missing will prevent deadlocks!
 Different avoidance strategies (which?)
32
Issues with Spinlocks

Spin-locking is very wasteful
 The spinning thread occupies resources
 Potentially the PE where the waiting thread wants to run  requires
context switch!

Context switches due to
 Expiration of time-slices (forced)
 Yielding the CPU
33
What is this?
34
Why is the 1997 Mars Rover in our lecture?

It landed, received program, and worked … until it spuriously
rebooted!
  watchdog

Scenario (vxWorks RT OS):
 Single CPU
 Two threads A,B sharing common bus, using locks
 (independent) thread C wrote data to flash
 Priority: ACB (A highest, B lowest)
 Thread C would run into a lifelock (infinite loop)
 Thread B was preempted by C while holding lock
 Thread A got stuck at lock 
[http://research.microsoft.com/en-us/um/people/mbj/Mars_Pathfinder/Authoritative_Account.html]
35
Priority Inversion


If busy-waiting thread has higher priority than thread holding lock ⇒
no progress!
Can be fixed with the help of the OS
 E.g., mutex priority inheritance (temporarily boost priority of task in CR to
highest priority among waiting tasks)
36
Condition Variables

Allow threads to yield CPU and leave the OS run queue
 Other threads can get them back on the queue!

cond_wait(cond, lock) – yield and go to sleep

cond_signal(cond) – wake up sleeping threads

Wait and signal are OS calls
 Often expensive, which one is more expensive?
Wait, because it has to perform a full context switch
37
Condition Variable Semantics

Hoare-style:
 Signaler passes lock to waiter, signaler suspended
 Waiter runs immediately
 Waiter passes lock back to signaler if it leaves critical section or if it waits
again

Mesa-style (most used):
 Signaler keeps lock
 Waiter simply put on run queue
 Needs to acquire lock, may wait again
38
When to Spin and When to Block?

Spinning consumes CPU cycles but is cheap
 “Steals” CPU from other threads

Blocking has high one-time cost and is then free
 Often hundreds of cycles (trap, save TCB …)
 Wakeup is also expensive (latency)
Also cache-pollution

Strategy:
 Poll for a while and then block
39
When to Spin and When to Block?

What is a “while”?

Optimal time depends on the future
 When will the active thread leave the CR?
 Can compute optimal offline schedule
 Actual problem is an online problem

Competitive algorithms
 An algorithm is c-competitive if for a sequence of actions x and a constant
a holds:
C(x) ≤ c*Copt(x) + a
 What would a good spinning algorithm look like and what is the
competitiveness?
40